Question

我正在将NLopt AUGLAG与COBYLA一起用作本地最小化函数。为什么会出现“舍入受限”错误？如何解决此错误？

这是我尝试过的事情：删除所有约束，但它仍然返回RoundoffLimited错误。查看优化程序的值和出错之前的目标值，但是对于最小化问题，目标函数值是无穷大。我也尝试过相对的容忍度，但仍然没有变化。

因为我使用的是AUGLAG，所以我认为不需要定义梯度，因为该算法是无梯度的。这是我的代码的一般结构：

# setting global variables to keep track of optimizers in case or error
previous_args = None
current_args = None

def myfunc(x, grad):
  global previous_args
  global current_args

  previous_args = current_args
  current_args = x

  occ = 0.5/x
  return sqrt(np.sum((target - occ)**2)/len(x))

def myconstraint(x, grad, a, b):
    return b - np.sum(x*a)

No = number of optimization variables I have
opt = nlopt.opt(nlopt.AUGLAG, No)

# Set local optimizer for AUGLAG
local_opt = nlopt.opt(nlopt.LN_COBYLA,No)
opt.set_local_optimizer(local_opt)

# Lower bounds for the variables
opt.set_lower_bounds(np.zeros((No)))

# Set objective fn
opt.set_min_objective(myfunc)

# Add constraints, tot_rev is a number
opt.add_inequality_constraint(lambda x,grad: myconstraint(x,grad,2/x,tot_rev), 1e-8)

try:
    x = opt.optimize(np.ones((No)))
except nlopt.RoundoffLimited:
    x = previous_args

# The exact error I get on x = opt.optimize is
# RoundoffLimited: NLopt roundoff-limited if I don't include the exception

# Set relative tolerance on optimization parameters
opt.set_xtol_rel(1e-3)

NLopt AUGLAG的舍入受限错误-发生此错误时如何获得合理的值，为什么会发生？

0 个答案: